CARLADrive: A Synthetic Multimodal Perception Dataset for Autonomous Driving

CARLADrive is a large-scale synthetic multimodal perception dataset for autonomous driving, built on CARLA Leaderboard 2.0 routes with a nuScenes-inspired sensor suite: 6 RGB cameras (1600x900 px, 360° coverage), 1 LiDAR (64 channels, 10 Hz, 80 m range), 5 radars (80 m range), 1 IMU/GNSS unit, 1 instance-segmentation camera, and 1 semantic LiDAR. Data is synchronized at 2 Hz and collected using PDM-Lite, a privileged rule-based expert agent that deterministically completes complex urban routes (unprotected turns, merging, yielding to emergency vehicles, etc.) across downtowns, neighborhoods, and highways The dataset provides over 1.6 million annotated 3D bounding boxes within a 60-meter perception range, covering the classes car, walker, bicycle, stop_sign, traffic_light, and static_trafficwarning, plus ego dynamics, future trajectories, navigation/intention labels, and dense semantic LiDAR point clouds.

It is released in two splits: - Mini: 4 training routes + 1 validation route, ~12,000 training / 4,000 validation samples, ~80 GB. - Standard: 30 training routes + 5 validation routes, ~82,000 training / 27,000 validation samples, spanning ~260 km and more than 11 hours of driving, ~400 GB Data is formatted into a KITTI-like structure (camera images, LiDAR/radar point clouds, 2D/3D labels, calibration files) via the accompanying format_dataset.py script.